Successful Distributed RL Training on Mac-based Infrastructure

Pluralis Research succeeded in generating RL (reinforcement learning) rollouts using 14 consumer-grade Macs distributed across 4 countries, relying only on the regular internet and Cloudflare R2 without any data center connectivity.

The key technical features are as follows:

• MLX-based int8 inference: Each Mac performs int8 quantized inference via MLX, while a single B200 GPU handles bf16 gradient updates.
• PULSE framework: Instead of transmitting full checkpoints during model updates, only int8 weight deltas are sent, dramatically reducing data transfer from 9GB to 82MB.
• DPPO-style probability gating: Tokens with large probability deviation between the rollout model and the trainer model are filtered out to prevent off-policy issues.

As a result of the experiment, on the PaperSearchQA benchmark, pass@1 performance improved dramatically from 29% to 63%, and tool use (search rate) improved from 22% to 84%. This provides a technical foundation for distributing the inference and training of large-scale models onto consumer hardware in the future.